AI UX: Stori’s on-device LLM for contextual microcopy cuts support tickets by 35%

AI · 6 min read

Stori faced high volumes of rookie user questions around export options, copyright prompts, and image attribution. Rather than route every query to a cloud LLM, the team evaluated a lightweight on-device LLM to provide immediate, contextual microcopy tailored to the current UI state. The goals were low latency, privacy, and dynamic messaging that could be A/B tested quickly.

They selected a 3B-parameter distilled model optimized for mobile inference, integrated through a local inference runtime, and built a microcopy engine that sends only the UI state and export metadata to the model. UX choices included short contextual hints that appear inline, a one-tap elaboration panel, and a concise footer explaining privacy implications of on-device generation. Designers also created a content governance layer to surface safe, pre-approved phrasing for legal-sensitive flows.

In an experiment across 10,000 users, the contextual microcopy reduced export-related support tickets by 35%, decreased export abandonment by 14%, and improved Net Promoter Score for new users. Latency stayed under 120ms on modern devices, and user trust improved because the messages did not require sending content to the cloud. The trade-off required additional QA for edge-case prompts and a maintenance pipeline to update the microcopy model and governance policies.